Aiming at the problem that the bit error rate (BER) of asymmetrically clipped optical orthogonal frequency division multiplexing (ACO-OFDM) space optical communication system is significantly affected by different turbulence intensities, the deep learning technique is proposed to the polarization code decoding in ACO-OFDM space optical communication system. Moreover, this system realizes the polarization code decoding and signal demodulation without frequency conduction with superior performance and robustness compared with the performance of traditional decoder. Simulations under different turbulence intensities as well as different mapping orders show that the convolutional neural network (CNN) decoder trained under weak-medium-strong turbulence atmospheric channels achieves a performance improvement of about 102 compared to the conventional decoder at 4-quadrature amplitude modulation (4QAM), and the BERs for both 16QAM and 64QAM are in between those of the conventional decoder.